JMIR Publications

ABSTRACT

Background:

Objective functional assessment of musculoskeletal (MSK) conditions remains a persistent clinical challenge, with current approaches relying on subjective physical examination and patient-reported outcome measures that lack consistency for scalable monitoring.

Objective:

This proof-of-concept study evaluated the feasibility of an extended reality (XR)-based framework for structured inertial measurement unit (IMU) kinematic data collection, automated machine learning (ML) classification of shoulder pathology, and clinically interpretable feature importance analysis.

Methods:

Six functional tasks derived from the Disabilities of the Arm, Shoulder, and Hand (DASH) instrument were implemented as goal-directed XR simulations. Forty patients with shoulder conditions and 20 healthy controls performed these tasks while 6-degree-of-freedom (6-DoF) kinematic data were captured from the XR headset and controllers at 50 Hz. Three model architectures were evaluated: recurrent neural networks (RNNs), convolutional neural networks (CNNs), and Transformers, across four clinical classification paradigms of increasing diagnostic specificity: patients versus controls, rotator cuff tears (RCTs) versus controls, other shoulder conditions versus controls, and RCTs versus other conditions. All models were assessed for classification performance and clinical explainability through feature importance analysis.

Results:

RNNs demonstrated the most consistent performance across tasks and paradigms (mean BA: 0.753, SD: 0.069), with peak performance on Jar Opening for RCTs versus controls (mean BA: 0.86, AUC: 0.82). Jar Opening, Back Washing, and Cutting yielded the highest discriminability across models. Paradigm 4 (RCTs versus other shoulder conditions) yielded the lowest classification performance across all models (BA: 0.49-0.71), with several models returning AUC values below 0.50. Head compensation emerged as the most important feature.

Conclusions:

Clinically grounded XR tasks produce structured, model-ready IMU data sufficient for shoulder pathology detection, demonstrating framework feasibility for objective MSK assessment. Head compensation was identified as a candidate kinematic marker of shoulder dysfunction. Differential diagnosis between pathology subgroups remains an open challenge, motivating task, and model designs tailored to specific clinical contrasts.